A Pilot Study of Learning from Examples to Translate Better

This project will explore advanced technologies and practical methodology to implement an online machine translation (MT) system that can learn to translate better. By providing an online translation service with a bilingual editor for manual post-editing, the system acquires translation knowledge from translators to enrich its example base and language models. The more knowledge it acquires, the better it translates. The baseline learning strategy is translation memory (TM): all translated texts (e.g., words, phrases, chunks) are memorized and need not be translated again. With adequate data of this kind the fundamental problem of empirical MT can be investigated: given a set of translated text chunks, how do we infer the optimal translation for a source sentence? This project proposes to divide this problem into three subproblems and tackle them with three statistical models: (1) a source language model to decompose the source sentence into an optimal sequence of chunks, (2) a translation model to select the best existing translation for each chunk, and (3) a target language model to recombine the translated chunks into a well-formed target sentence. A unique feature of this system is that it adapts its translation towards translators' expertise via learning. The research will focus on machine translation for the language pair of English and Chinese.